Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ViTmiX: Vision Transformer Explainability Augmented by Mixed Visualization Methods

Eduard Hogea, Darian M. Onchis, Ana Coporan, Adina Magda Florea, Codruta Istin

TL;DR

This paper addresses the explainability challenges of Vision Transformers by proposing ViTmiX, a hybrid approach that blends multiple heatmap-based XAI methods. It analyzes several explainability techniques, including LRP, CAM, Rollout, and Saliency, and systematically evaluates pairwise and triple-method combinations, using geometric mean as a fusion operator. Experiments on Pascal VOC and the PH2 dataset show that two-way mixes, particularly LRP+Rollout with geometric mean, yield the strongest segmentation and localization performance, often surpassing individual methods. Additionally, the study introduces a post-hoc explainability measure based on the Pigeonhole principle to quantify explainability gains, and confirms that hybrid explanations provide more faithful object localization in ViTs.

Abstract

Recent advancements in Vision Transformers (ViT) have demonstrated exceptional results in various visual recognition tasks, owing to their ability to capture long-range dependencies in images through self-attention mechanisms. However, the complex nature of ViT models requires robust explainability methods to unveil their decision-making processes. Explainable Artificial Intelligence (XAI) plays a crucial role in improving model transparency and trustworthiness by providing insights into model predictions. Current approaches to ViT explainability, based on visualization techniques such as Layer-wise Relevance Propagation (LRP) and gradient-based methods, have shown promising but sometimes limited results. In this study, we explore a hybrid approach that mixes multiple explainability techniques to overcome these limitations and enhance the interpretability of ViT models. Our experiments reveal that this hybrid approach significantly improves the interpretability of ViT models compared to individual methods. We also introduce modifications to existing techniques, such as using geometric mean for mixing, which demonstrates notable results in object segmentation tasks. To quantify the explainability gain, we introduced a novel post-hoc explainability measure by applying the Pigeonhole principle. These findings underscore the importance of refining and optimizing explainability methods for ViT models, paving the way to reliable XAI-based segmentations.

ViTmiX: Vision Transformer Explainability Augmented by Mixed Visualization Methods

TL;DR

This paper addresses the explainability challenges of Vision Transformers by proposing ViTmiX, a hybrid approach that blends multiple heatmap-based XAI methods. It analyzes several explainability techniques, including LRP, CAM, Rollout, and Saliency, and systematically evaluates pairwise and triple-method combinations, using geometric mean as a fusion operator. Experiments on Pascal VOC and the PH2 dataset show that two-way mixes, particularly LRP+Rollout with geometric mean, yield the strongest segmentation and localization performance, often surpassing individual methods. Additionally, the study introduces a post-hoc explainability measure based on the Pigeonhole principle to quantify explainability gains, and confirms that hybrid explanations provide more faithful object localization in ViTs.

Abstract

Recent advancements in Vision Transformers (ViT) have demonstrated exceptional results in various visual recognition tasks, owing to their ability to capture long-range dependencies in images through self-attention mechanisms. However, the complex nature of ViT models requires robust explainability methods to unveil their decision-making processes. Explainable Artificial Intelligence (XAI) plays a crucial role in improving model transparency and trustworthiness by providing insights into model predictions. Current approaches to ViT explainability, based on visualization techniques such as Layer-wise Relevance Propagation (LRP) and gradient-based methods, have shown promising but sometimes limited results. In this study, we explore a hybrid approach that mixes multiple explainability techniques to overcome these limitations and enhance the interpretability of ViT models. Our experiments reveal that this hybrid approach significantly improves the interpretability of ViT models compared to individual methods. We also introduce modifications to existing techniques, such as using geometric mean for mixing, which demonstrates notable results in object segmentation tasks. To quantify the explainability gain, we introduced a novel post-hoc explainability measure by applying the Pigeonhole principle. These findings underscore the importance of refining and optimizing explainability methods for ViT models, paving the way to reliable XAI-based segmentations.

Paper Structure

This paper contains 9 sections, 3 equations, 7 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Evaluation Metrics
  5. Experimental assessment
  6. Model Performance
  7. Pascal VOC
  8. PH2 Dataset
  9. Conclusions

Figures (7)

  • Figure 1: ViTmiX: Complementarity of explainable AI methods. Each feature attribution map represents the relevance or importance scores calculated by the respective methods applied to a visual transformer model. Each cell in the heatmap visualization corresponds to the relevance or importance of a particular feature (e.g., a pixel in an image). The colors (Blues, Reds, Greens, Oranges, Purples) are used to visually differentiate between the methods (CAM, LRP, Rollout Attention, Saliency Map) and their mixed features attribution map.
  • Figure 2: Heatmaps of proposed visualization techniques — Multiplication
  • Figure 3: Heatmaps of proposed visualization techniques — Geometric Mean
  • Figure 4: Heatmaps of existing visualization techniques
  • Figure 5: Analysis of the results of the LRP+CAM 2WAY method. While the generated mask appears reasonable at first glance, the ground truth mask in this case seems to be incorrect.
  • ...and 2 more figures